The "sendfile_max_chunk" directive is important to prevent worker
monopolization by fast connections. The 2m value implies maximum 200ms
delay with 100 Mbps links, 20ms delay with 1 Gbps links, and 2ms on
10 Gbps links. It also seems to be a good value for disks.
Previously, connections to upstream servers used sendfile() if it was
enabled, but never honored sendfile_max_chunk. This might result
in worker monopolization for a long time if large request bodies
are allowed.
On Linux starting with 2.6.16, sendfile() silently limits all operations
to MAX_RW_COUNT, defined as (INT_MAX & PAGE_MASK). This incorrectly
triggered the interrupt check, and resulted in 0-sized writev() on the
next loop iteration.
Fix is to make sure the limit is always checked, so we will return from
the loop if the limit is already reached even if number of bytes sent is
not exactly equal to the number of bytes we've tried to send.
Previously, it was checked that sendfile_max_chunk was enabled and
almost whole sendfile_max_chunk was sent (see e67ef50c3176), to avoid
delaying connections where sendfile_max_chunk wasn't reached (for example,
when sending responses smaller than sendfile_max_chunk). Now we instead
check if there are unsent data, and the connection is still ready for writing.
Additionally we also check c->write->delayed to ignore connections already
delayed by limit_rate.
This approach is believed to be more robust, and correctly handles
not only sendfile_max_chunk, but also internal limits of c->send_chain(),
such as sendfile() maximum supported length (ticket #1870).
Previously, 1 millisecond delay was used instead. In certain edge cases
this might result in noticeable performance degradation though, notably on
Linux with typical CONFIG_HZ=250 (so 1ms delay becomes 4ms),
sendfile_max_chunk 2m, and link speed above 2.5 Gbps.
Using posted next events removes the artificial delay and makes processing
fast in all cases.
The directive enables including all frames from start time to the most recent
key frame in the result. Those frames are removed from presentation timeline
using mp4 edit lists.
Edit lists are currently supported by popular players and browsers such as
Chrome, Safari, QuickTime and ffmpeg. Among those not supporting them properly
is Firefox[1].
Based on a patch by Tracey Jaquith, Internet Archive.
[1] https://bugzilla.mozilla.org/show_bug.cgi?id=1735300
The function updates the duration field of mdhd atom. Previously it was
updated in ngx_http_mp4_read_mdhd_atom(). The change makes it possible to
alter track duration as a result of processing track frames.
Hash initialization ignores elements with key.data set to NULL.
Nevertheless, the initial hash bucket size check didn't skip them,
resulting in unnecessary restrictions on, for example, variables with
long names and with the NGX_HTTP_VARIABLE_NOHASH flag.
Fix is to update the initial hash bucket size check to skip elements
with key.data set to NULL, similarly to how it is done in other parts
of the code.
Requires OpenSSL 3.0 compiled with "enable-ktls" option. Further, KTLS
needs to be enabled in kernel, and in OpenSSL, either via OpenSSL
configuration file or with "ssl_conf_command Options KTLS;" in nginx
configuration.
On FreeBSD, kernel TLS is available starting with FreeBSD 13.0, and
can be enabled with "sysctl kern.ipc.tls.enable=1" and "kldload ktls_ocf"
to load a software backend, see man ktls(4) for details.
On Linux, kernel TLS is available starting with kernel 4.13 (at least 5.2
is recommended), and needs kernel compiled with CONFIG_TLS=y (with
CONFIG_TLS=m, which is used at least on Ubuntu 21.04 by default,
the tls module needs to be loaded with "modprobe tls").
While clock_gettime(CLOCK_MONOTONIC_COARSE) is faster than
clock_gettime(CLOCK_MONOTONIC), the latter is fast enough on Linux for
practical usage, and the difference is negligible compared to other costs
at each event loop iteration. On the other hand, CLOCK_MONOTONIC_COARSE
causes various issues with typical CONFIG_HZ=250, notably very inaccurate
limit_rate handling in some edge cases (ticket #1678) and negative difference
between $request_time and $upstream_response_time (ticket #1965).
This is a recommended behavior by RFC 7301 and is useful
for mitigation of protocol confusion attacks [1].
To avoid possible negative effects, list of supported protocols
was extended to include all possible HTTP protocol ALPN IDs
registered by IANA [2], i.e. "http/1.0" and "http/0.9".
[1] https://alpaca-attack.com/
[2] https://www.iana.org/assignments/tls-extensiontype-values/
In b87b7092cedb (nginx 1.21.1), logging level of "upstream sent invalid
header" errors was accidentally changed to "info". This change restores
the "error" level, which is a proper logging level for upstream-side
errors.
The u->keepalive flag is initialized early if the response has no body
(or an empty body), and needs to be reset if there are any extra data,
similarly to how it is done in ngx_http_proxy_copy_filter(). Missed
in 83c4622053b0.
The "proxy_half_close" directive enables handling of TCP half close. If
enabled, connection to proxied server is kept open until both read ends get
EOF. Write end shutdown is properly transmitted via proxy.
Do this only when the entire request body is empty and
r->request_body_in_file_only is set.
The issue manifested itself with missing warning "a client request body is
buffered to a temporary file" when the entire rb->buf is full and all buffers
are delayed by a filter.
The request body filter chain is no longer called after processing
a DATA frame. Instead, we now post a read event to do this. This
ensures that multiple small DATA frames read during the same event loop
iteration are coalesced together, resulting in much faster processing.
Since rb->buf can now contain unprocessed data, window update is no
longer sent in ngx_http_v2_state_read_data() in case of flow control
being used due to filter buffering. Instead, window will be updated
by ngx_http_v2_read_client_request_body_handler() in the posted read
event.
Following rb->filter_need_buffering changes, request body reading is
only finished after the filter chain is called and rb->last_saved is set.
As such, with r->request_body_no_buffering, timer on fc->read is no
longer removed when the last part of the body is received, potentially
resulting in incorrect behaviour.
The fix is to call ngx_http_v2_process_request_body() from the
ngx_http_v2_read_unbuffered_request_body() function instead of
directly calling ngx_http_v2_filter_request_body(), so the timer
is properly removed.
In the body read handler, the window was incorrectly calculated
based on the full buffer size instead of the amount of free space
in the buffer. If the request body is buffered by a filter, and
the buffer is not empty after the read event is generated by the
filter to resume request body processing, this could result in
"http2 negative window update" alerts.
Further, in the body ready handler and in ngx_http_v2_state_read_data()
the buffer wasn't cleared when the data were already written to disk,
so the client might stuck without window updates.
If a filter wants to buffer the request body during reading (for
example, to check an external scanner), it can now do so. To make
it possible, the code now checks rb->last_saved (introduced in the
previous change) along with rb->rest == 0.
Since in HTTP/2 this requires flow control to avoid overflowing the
request body buffer, so filters which need buffering have to set
the rb->filter_need_buffering flag on the first filter call. (Note
that each filter is expected to call the next filter, so all filters
will be able set the flag if needed.)
It indicates that the last buffer was received by the save filter,
and can be used to check this at higher levels. To be used in the
following changes.
If due to an error ngx_http_request_body_save_filter() is called
more than once with rb->rest == 0, this used to result in a segmentation
fault. Added an alert to catch such errors, just in case.
Previously, fully preread unbuffered requests larger than client body
buffer size were saved to disk, despite the fact that "unbuffered" is
expected to imply no disk buffering.
The save body filter saves the request body to disk once the buffer is full.
Yet in HTTP/2 this might happen even if there is no need to save anything
to disk, notably when content length is known and the END_STREAM flag is
sent in a separate empty DATA frame. Workaround is to provide additional
byte in the buffer, so saving the request body won't be triggered.
This fixes unexpected request body disk buffering in HTTP/2 observed after
the previous change when content length is known and the END_STREAM flag
is sent in a separate empty DATA frame.
In particular, now the code always uses a buffer limited by
client_body_buffer_size. At the cost of an additional copy it
ensures that small DATA frames are not directly mapped to small
write() syscalls, but rather buffered in memory before writing.
Further, requests without Content-Length are no longer forced
to use temporary files.
With SSL it is possible that an established connection is ready for
reading after the handshake. Further, events might be already disabled
in case of level-triggered event methods. If this happens and
ngx_http_upstream_send_request() blocks waiting for some data from
the upstream, such as flow control in case of gRPC, the connection
will time out due to no read events on the upstream connection.
Fix is to explicitly check the c->read->ready flag if sending request
blocks and post a read event if it is set.
Note that while it is possible to modify ngx_ssl_handshake() to keep
read events active, this won't completely resolve the issue, since
there can be data already received during the SSL handshake
(see 573bd30e46b4).
This adds new Auth-SSL-Protocol and Auth-SSL-Cipher headers to
the mail proxy auth protocol when SSL is enabled.
This can be useful for detecting users using older clients that
negotiate old ciphers when you want to upgrade to newer
TLS versions of remove suppport for old and insecure ciphers.
You can use your auth backend to notify these users before the
upgrade that they either need to upgrade their client software
or contact your support team to work out an upgrade path.
To load old/weak server or client certificates it might be needed to adjust
the security level, as introduced in OpenSSL 1.1.0. This change ensures that
ciphers are set before loading the certificates, so security level changes
via the cipher string apply to certificate loading.
Export ciphers are forbidden to negotiate in TLS 1.1 and later protocol modes.
They are disabled since OpenSSL 1.0.2g by default unless explicitly configured
with "enable-weak-ssl-ciphers", and completely removed in OpenSSL 1.1.0.
A new behaviour was introduced in OpenSSL 1.1.1e, when a peer does not send
close_notify before closing the connection. Previously, it was to return
SSL_ERROR_SYSCALL with errno 0, known since at least OpenSSL 0.9.7, and is
handled gracefully in nginx. Now it returns SSL_ERROR_SSL with a distinct
reason SSL_R_UNEXPECTED_EOF_WHILE_READING ("unexpected eof while reading").
This leads to critical errors seen in nginx within various routines such as
SSL_do_handshake(), SSL_read(), SSL_shutdown(). The behaviour was restored
in OpenSSL 1.1.1f, but presents in OpenSSL 3.0 by default.
Use of the SSL_OP_IGNORE_UNEXPECTED_EOF option added in OpenSSL 3.0 allows
to set a compatible behaviour to return SSL_ERROR_ZERO_RETURN:
https://git.openssl.org/?p=openssl.git;a=commitdiff;h=09b90e0
See for additional details: https://github.com/openssl/openssl/issues/11381
The OPENSSL_SUPPRESS_DEPRECATED macro is used to suppress deprecation warnings.
This covers Session Tickets keys, SSL Engine, DH low level API for DHE ciphers.
Unlike OPENSSL_API_COMPAT, it works well with OpenSSL built with no-deprecated.
In particular, it doesn't unhide various macros in OpenSSL includes, which are
meant to be hidden under OPENSSL_NO_DEPRECATED.
The only consumer is a callback function for SSL_CTX_set_tmp_rsa_callback()
deprecated in OpenSSL 1.1.0. Now the function is conditionally compiled too.
The latest HTTP/1.1 draft describes Transfer-Encoding in HTTP/1.0 as having
potentially faulty message framing as that could have been forwarded without
handling of the chunked encoding, and forbids processing subsequest requests
over that connection: https://github.com/httpwg/http-core/issues/879.
While handling of such requests is permitted, the most secure approach seems
to reject them.
The c->read->ready and c->write->ready flags might be reset during
the handshake, and not set again if the handshake was finished on
the other event. At the same time, some data might be read from
the socket during the handshake, so missing c->read->ready flag might
result in a connection hang, for example, when waiting for an SMTP
greeting (which was already received during the handshake).
Found by Sergey Kandaurov.
The ngx_wsasend_chain() and ngx_wsarecv_chain() functions were
modified to use only preallocated memory, and the number of
preallocated wsabufs was increased to 64.
In d1bde5c3c5d2, the number of preallocated iovec's for ngx_readv_chain()
was increased. Still, in some setups, the function might allocate memory
for iovec's from a connection pool, which is only freed when closing the
connection.
The ngx_readv_chain() function was modified to use only preallocated
memory, similarly to the ngx_writev_chain() change in 8e903522c17a.
Control characters (0x00-0x1f, 0x7f) and space are not expected to appear
in the Host header. Requests with such characters in the Host header are
now unconditionally rejected.
In 71edd9192f24 logging of invalid headers which were rejected with the
NGX_HTTP_PARSE_INVALID_HEADER error was restricted to just the "client
sent invalid header line" message, without any attempts to log the header
itself.
This patch returns logging of the header up to the invalid character and
the character itself. The r->header_end pointer is now properly set
in all cases to make logging possible.
The same logging is also introduced when parsing headers from upstream
servers.
Control characters (0x00-0x1f, 0x7f), space, and colon were never allowed in
header names. The only somewhat valid use is header continuation which nginx
never supported and which is explicitly obsolete by RFC 7230.
Previously, such headers were considered invalid and were ignored by default
(as per ignore_invalid_headers directive). With this change, such headers
are unconditionally rejected.
It is expected to make nginx more resilient to various attacks, in particular,
with ignore_invalid_headers switched off (which is inherently unsecure, though
nevertheless sometimes used in the wild).
Control characters (0x00-0x1f, 0x7f) were never allowed in URIs, and must
be percent-encoded by clients. Further, these are not believed to appear
in practice. On the other hand, passing such characters might make various
attacks possible or easier, despite the fact that currently allowed control
characters are not significant for HTTP request parsing.
From now on, requests with spaces in URIs are immediately rejected rather
than allowed. Spaces were allowed in 31e9677b15a1 (0.8.41) to handle bad
clients. It is believed that now this behaviour causes more harm than
good.
Per RFC 3986 only the following characters are allowed in URIs unescaped:
unreserved = ALPHA / DIGIT / "-" / "." / "_" / "~"
gen-delims = ":" / "/" / "?" / "#" / "[" / "]" / "@"
sub-delims = "!" / "$" / "&" / "'" / "(" / ")"
/ "*" / "+" / "," / ";" / "="
And "%" can appear as a part of escaping itself. The following
characters are not allowed and need to be escaped: %00-%1F, %7F-%FF,
" ", """, "<", ">", "\", "^", "`", "{", "|", "}".
Not escaping ">" is known to cause problems at least with MS Exchange (see
http://nginx.org/pipermail/nginx-ru/2010-January/031261.html) and in
Tomcat (ticket #2191).
The patch adds escaping of the following chars in all URI parts: """, "<",
">", "\", "^", "`", "{", "|", "}". Note that comments are mostly preserved
to outline important characters being escaped.
HTTP clients are not allowed to generate such requests since Transfer-Encoding
introduction in RFC 2068, and they are not expected to appear in practice
except in attempts to perform a request smuggling attack. While handling of
such requests is strictly defined, the most secure approach seems to reject
them.
No valid CONNECT requests are expected to appear within nginx, since it
is not a forward proxy. Further, request line parsing will reject
proper CONNECT requests anyway, since we don't allow authority-form of
request-target. On the other hand, RFC 7230 specifies separate message
length rules for CONNECT which we don't support, so make sure to always
reject CONNECTs to avoid potential abuse.
Previously, TRACE requests were rejected before parsing Transfer-Encoding.
This is not important since keepalive is not enabled at this point anyway,
though rejecting such requests after properly parsing other headers is
less likely to cause issues in case of further code changes.
After 2096b21fcd10, a single RST_STREAM(NO_ERROR) may not result in an error.
This change removes several unnecessary ctx->type checks for such a case.
Recent fixes to SSL shutdown with lingering close (554c6ae25ffc, 1.19.5)
broke logging of SSL variables. To make sure logging of SSL variables
works properly, avoid freeing c->ssl when doing an SSL shutdown before
lingering close.
Reported by Reinis Rozitis
(http://mailman.nginx.org/pipermail/nginx/2021-May/060670.html).
Instead of calling SSL_free() with each return point, introduced a single
place where cleanup happens. As a positive side effect, this fixes two
potential memory leaks on ngx_handle_read_event() and ngx_handle_write_event()
errors where there were no SSL_free() calls (though unlikely practical,
as errors there are only expected to happen due to bugs or kernel issues).
On Linux, SO_REUSEADDR allows completely duplicate UDP sockets, so using
SO_REUSEADDR when testing configuration results in packets being dropped
if there is an existing traffic on the sockets being tested (ticket #2187).
While dropped packets are expected with UDP, it is better to avoid this
when possible.
With this change, SO_REUSEADDR is no longer set on datagram sockets when
testing configuration.
Since we anyway do not set SO_REUSEPORT when testing configuration
(see ecb5cd305b06), trying to open additional sockets does not make much
sense, as all these additional sockets are expected to result in EADDRINUSE
errors from bind(). On the other hand, there are reports that trying
to open these sockets takes significant time under load: total configuration
testing time greater than 15s was observed in ticket #2188, compared to less
than 1s without load.
With this change, no additional sockets are opened during testing
configuration.
Since nginx always uses exactly one entry in the question section of
a DNS query, and never uses compression pointers in this entry, parsing
of a DNS response in ngx_resolver_process_response() does not expect
compression pointers to appear in the question section of the DNS
response. Indeed, compression pointers in the first name of a DNS response
hardly make sense, do not seem to be allowed by RFC 1035 (which says
"a pointer to a prior occurance of the same name", note "prior"), and
were never observed in practice.
Added an explicit check to ngx_resolver_process_response()'s parsing
of the question section to properly report an error if compression pointers
nevertheless appear in the question section.
Instead of checking on each label if we need to place a dot or not,
now it always adds a dot after a label, and reduces the resulting
length afterwards.
Previously, anything with any of the two high bits set were interpreted
as compression pointers. This is incorrect, as RFC 1035 clearly states
that "The 10 and 01 combinations are reserved for future use". Further,
the 01 combination is actually allocated for EDNS extended label type
(see RFC 2671 and RFC 6891), not really used though.
Fix is to reject unrecognized label types rather than misinterpreting
them as compression pointers.
It is believed to be harmless, and in the worst case it uses some
uninitialized memory as a part of the compression pointer length,
eventually leading to the "name is out of DNS response" error.
When variables are used in ssl_certificate or ssl_certificate_key, a request
is created in the certificate callback to evaluate the variables, and then
freed. Freeing it, however, updates c->log->action to "closing request",
resulting in confusing error messages like "client timed out ... while
closing request" when a client times out during the SSL handshake.
Fix is to restore c->log->action after calling ngx_http_free_request().
Similarly to smtpd_hard_error_limit in Postfix and smtp_max_unknown_commands
in Exim, specifies the number of errors after which the connection is closed.
The change is mostly the same as the SMTP one (04e43d03e153 and 3f5d0af4e40a),
and ensures that nginx is able to properly handle or reject multiple IMAP
commands. The s->cmd field is not really used and set for consistency.
Non-synchronizing literals handling in invalid/unknown commands is limited,
so when a non-synchronizing literal is detected at the end of a discarded
line, the connection is closed.
Only "A-Za-z0-9-._" characters now allowed (which is stricter than what
RFC 3501 requires, but expected to be enough for all known clients),
and tags shouldn't be longer than 32 characters.
Previously, s->backslash was set if any of the arguments was a quoted
string with a backslash character. After successful command parsing
this resulted in all arguments being filtered to remove backslashes.
This is, however, incorrect, as backslashes should not be removed from
IMAP literals. For example:
S: * OK IMAP4 ready
C: a01 login {9}
S: + OK
C: user\name "pass\"word"
S: * BAD internal server error
resulted in "Auth-User: username" instead of "Auth-User: user\name"
as it should.
Fix is to apply backslash filtering on per-argument basis during parsing.
As discussed in the previous change, s->arg_start handling in the "done"
labels of ngx_mail_pop3_parse_command(), ngx_mail_imap_parse_command(),
and ngx_mail_smtp_parse_command() is wrong: s->arg_start cannot be
set there, as it is handled and cleared on all code paths where the
"done" labels are reached. The relevant code is dead and now removed.
Previously, s->arg_start was left intact after invalid IMAP commands,
and this might result in an argument incorrectly added to the following
command. Similarly, s->backslash was left intact as well, leading
to unneeded backslash removal.
For example (LFs from the client are explicitly shown as "<LF>"):
S: * OK IMAP4 ready
C: a01 login "\<LF>
S: a01 BAD invalid command
C: a0000000000\2 authenticate <LF>
S: a00000000002 aBAD invalid command
The backslash followed by LF generates invalid command with s->arg_start
and s->backslash set, the following command incorrectly treats anything
from the old s->arg_start to the space after the command as an argument,
and removes the backslash from the tag. If there is no space, s->arg_end
will be NULL.
Both things seem to be harmless though. In particular:
- This can be used to provide an incorrect argument to a command without
arguments. The only command which seems to look at the single argument
is AUTHENTICATE, and it checks the argument length before trying to
access it.
- Backslash removal uses the "end" pointer, and stops due to "src < end"
condition instead of scanning all the process memory if s->arg_end is
NULL (and arg[0].len is huge).
- There should be no backslashes in unquoted strings.
An obvious fix is to clear s->arg_start and s->backslash on invalid commands,
similarly to how it is done in POP3 parsing (added in 810:e3aa8f305d21) and
SMTP parsing.
This, however, makes it clear that s->arg_start handling in the "done"
label is wrong: s->arg_start cannot be legitimately set there, as it
is expected to be cleared in all possible cases when the "done" label is
reached. The relevant code is dead and will be removed by the following
change.
The change is mostly the same as the SMTP one (04e43d03e153 and 3f5d0af4e40a),
and ensures that nginx is able to properly handle or reject multiple POP3
commands, as required by the PIPELINING capability (RFC 2449). The s->cmd
field is not really used and set for consistency.
There is no need to scan buffer from s->buffer->pos, as we already scanned
the buffer till "p" and wasn't able to find an LF.
There is no real need for this change in SMTP, since it is at most a
microoptimization of a non-common code path. Similar code in IMAP, however,
will have to start scanning from "p" to be correct, since there can be
newlines in IMAP literals.
Previously, if an invalid SMTP command was split between reads, nginx failed
to wait for LF before returning an error, and interpreted the rest of the
command received later as a separate command.
The sw_invalid state in ngx_mail_smtp_parse_command(), introduced in
04e43d03e153, did not work, since ngx_mail_smtp_auth_state() clears
s->state when returning an error due to NGX_MAIL_PARSE_INVALID_COMMAND.
And not clearing s->state will introduce another problem: the rest
of the command would trigger duplicate error when rest of the command is
received.
Fix is to return NGX_AGAIN from ngx_mail_smtp_parse_command() until full
command is received.
Previously, if there were some pipelined SMTP data in the buffer when
a proxied connection with the backend was established, nginx called
ngx_mail_proxy_handler() to send these data, and not tried to send the
response to the last command. In most cases, this response was later sent
along with the response to the pipelined command, but if for some reason
client decides to wait for the response before finishing the next command
this might result in a connection hang.
Fix is to always call ngx_mail_proxy_handler() to send the response, and
additionally post an event to send the pipelined data if needed.
With this change, it is now possible to use ngx_conf_merge_ptr_value()
to merge complex values. This change follows much earlier changes in
ngx_conf_merge_ptr_value() and ngx_conf_set_str_array_slot()
in 1452:cd586e963db0 (0.6.10) and 1701:40d004d95d88 (0.6.22), and the
change in ngx_conf_set_keyval_slot() (7728:485dba3e2a01, 1.19.4).
To preserve compatibility with existing 3rd party modules, both NULL
and NGX_CONF_UNSET_PTR are accepted for now.
Due to structure's alignment, some uninitialized memory contents may have
been passed between processes.
Zeroing was removed in 0215ec9aaa8a.
Reported by Johnny Wang.
With SMTP pipelining, ngx_mail_read_command() can be called with s->buffer
without any space available, to parse additional commands received to the
buffer on previous calls. Previously, this resulted in recv() being called
with zero length, resulting in zero being returned, which was interpreted
as a connection close by the client, so nginx silently closed connection.
Fix is to avoid calling c->recv() if there is no free space in the buffer,
but continue parsing of the already received commands.
It turns out no browsers implement HTTP/2 GOAWAY handling properly, and
large enough number of resources on a page results in failures to load
some resources. In particular, Chrome seems to experience errors if
loading of all resources requires more than 1 connection (while it
is usually able to retry requests at least once, even with 2 connections
there are occasional failures for some reason), Safari if loading requires
more than 3 connections, and Firefox if loading requires more than 10
connections (can be configured with network.http.request.max-attempts,
defaults to 10).
It does not seem to be possible to resolve this on nginx side, even strict
limiting of maximum concurrency does not help, and loading issues seems to
be triggered by merely queueing of a request for a particular connection.
The only available mitigation seems to use higher keepalive_requests value.
The new default is 1000 and matches previously used default for
http2_max_requests. It is expected to be enough for 99.98% of the pages
(https://httparchive.org/reports/state-of-the-web?start=latest#reqTotal)
even in Chrome.
Similar to lingering_time, it limits total connection lifetime before
keepalive is switched off. The default is 1 hour, which is close to
the total maximum connection lifetime possible with default
keepalive_requests and keepalive_timeout.
Firefox uses several idle streams for PRIORITY frames[1], and
"http2_max_concurrent_streams 1;" results in "client sent too many
PRIORITY frames" errors when a connection is established by Firefox.
Fix is to relax the PRIORITY frames limit to use at least 100 as
the initial value (which is the recommended by the HTTP/2 protocol
minimum limit on the number of concurrent streams, so it is not
unreasonable for clients to assume that similar number of idle streams
can be used for prioritization).
[1] https://hg.mozilla.org/mozilla-central/file/32a9e6e145d6e3071c3993a20bb603a2f388722b/netwerk/protocol/http/Http2Stream.cpp#l1270
In current versions (all versions based on zlib 1.2.11, at least
since 2018) it no longer uses 64K hash and does not force window
bits to 13 if it is less than 13. That is, it needs just 16 bytes
more memory than normal zlib, so these bytes are simply added to
the normal size calculation.
In limit_req, auth_delay, and upstream code to check for broken
connections, tests for possible connection close by the client
did not work if the connection was already closed when relevant
event handler was set. This happened because there were no additional
events in case of edge-triggered event methods, and read events
were disabled in case of level-triggered ones.
Fix is to explicitly post a read event if the c->read->ready flag
is set.
For new data to be reported with eventport on Solaris,
ngx_handle_read_event() needs to be called after reading response
headers. To do so, ngx_http_upstream_process_non_buffered_upstream()
now called unconditionally if there are no prepread data. This
won't cause any read() syscalls as long as upstream connection
is not ready for reading (c->read->ready is not set), but will result
in proper handling of all events.
If we need to be notified about further events, ngx_handle_read_event()
needs to be called after a read event is processed. Without this,
an event can be removed from the kernel and won't be reported again,
notably when using oneshot event methods, such as eventport on Solaris.
While here, error handling is also added, similar to one present in
ngx_resolver_tcp_read(). This is not expected to make a difference
and mostly added for consistency.
If an attempt is made to delete an event which was already reported,
port_dissociate() returns an error. Fix is avoid doing anything if
ev->active is not set.
Possible alternative approach would be to avoid calling ngx_del_event()
at all if ev->active is not set. This approach, however, will require
something else to re-add the other event of the connection, since both
read and write events are dissociated if an event is reported on a file
descriptor. Currently ngx_eventport_del_event() re-associates write
event if called to delete read event, and vice versa.
If, at the start of an event loop iteration, there are any timers
in the past (including timers expiring now), the ngx_process_events()
function is called with zero timeout, and returns immediately even
if there are no events. But the following code only calls
ngx_event_expire_timers() if time actually changed, so this results
in nginx spinning in the event loop till current time changes.
While such timers are not expected to appear under normal conditions,
as all such timers should be removed on previous event loop iterations,
they still can appear due to bugs, zero timeouts set in the configuration
(if this is not explicitly handled by the code), or due to external
time changes on systems without clock_gettime(CLOCK_MONOTONIC).
Fix is to call ngx_event_expire_timers() unconditionally. Calling
it on each event loop iteration is not expected to be significant from
performance point of view, especially compared to a syscall in
ngx_process_events().
Without explicit handling, a zero timer was actually added, leading to
multiple unneeded syscalls. Further, sending GOAWAY frame early might
be beneficial for clients.
Reported by Sergey Kandaurov.
Unlike in 75e908236701, which added the logic to ngx_http_finalize_request(),
this change moves it to a more generic routine ngx_http_finalize_connection()
to cover cases when a request is finalized with NGX_DONE.
In particular, this fixes unwanted connection transition into the keepalive
state after receiving EOF while discarding request body. With edge-triggered
event methods that means the connection will last for extra seconds as set in
the keepalive_timeout directive.
The response size check introduced in 39501ce97e29 did not take into
account possible padding on DATA frames, resulting in incorrect
"upstream sent response body larger than indicated content length" errors
if upstream server used padding in responses with known length.
Fix is to check the actual size of response buffers produced by the code,
similarly to how it is done in other protocols, instead of checking
the size of DATA frames.
Reported at:
http://mailman.nginx.org/pipermail/nginx-devel/2021-March/013907.html
Activated with the "proxy_protocol" directive. Can be combined with
"listen ... proxy_protocol;" and "set_real_ip_from ...;" to pass
client address provided to nginx in the PROXY protocol header.
Activated with the "proxy_protocol" parameter of the "listen" directive.
Obtained information is passed to the auth_http script in Proxy-Protocol-Addr,
Proxy-Protocol-Port, Proxy-Protocol-Server-Addr, and Proxy-Protocol-Server-Port
headers.
Similarly to 40e8ce405859 in the stream module, this reduces the time
accept mutex is held. This also simplifies following changes to
introduce PROXY protocol support.
If we need to be notified about further events, ngx_handle_read_event()
needs to be called after a read event is processed. Without this,
an event can be removed from the kernel and won't be reported again,
notably when using oneshot event methods, such as eventport on Solaris.
For consistency, existing ngx_handle_read_event() call removed from
ngx_mail_read_command(), as this call only covers one of the code paths
where ngx_mail_read_command() returns NGX_AGAIN. Instead, appropriate
processing added to the callers, covering all code paths where NGX_AGAIN
is returned.
As long as a read event is blocked (ignored), ngx_handle_read_event()
needs to be called to make sure no further notifications will be
triggered when using level-triggered event methods, such as select() or
poll().
The "!rev->ready" test seems to be a typo, introduced in the original
commit (719:f30b1a75fd3b). The ngx_handle_write_event() code properly
tests for "rev->ready" instead.
Due to this typo, read events might be unexpectedly removed during
proxying after an event on the other part of the proxied connection.
Catched by mail proxying tests.
The strerrordesc_np() function, introduced in glibc 2.32, provides an
async-signal-safe way to obtain error messages. This makes it possible
to avoid copying error messages.
Previously, systems without sys_nerr (or _sys_nerr) were handled with an
assumption that errors start at 0 and continuous. This is, however, not
something POSIX requires, and not true on some platforms.
Notably, on Linux, where sys_nerr is no longer available for newly linked
binaries starting with glibc 2.32, there are gaps in error list, which
used to stop us from properly detecting maximum errno. Further, on
GNU/Hurd errors start at 0x40000001.
With this change, maximum errno detection is moved to the runtime code,
now able to ignore gaps, and also detects the first error if needed.
This fixes observed "Unknown error" messages as seen on Linux with
glibc 2.32 and on GNU/Hurd.
With this change, behaviour of HTTP/2 becomes even closer to HTTP/1.x,
and client_header_timeout instead of keepalive_timeout is used before
the first request is received.
This fixes HTTP/2 connections being closed even before the first request
if "keepalive_timeout 0;" was used in the configuration; the problem
appeared in f790816a0e87 (1.19.7).
Previously, PINGs and other frames extended possible keepalive time,
making it possible to keep an open HTTP/2 connection for a long time.
Now the connection is always closed as long as keepalive_timeout expires,
similarly to how it happens in HTTP/1.x.
Note that as a part of this change, incomplete frames are no longer
trigger a separate timeout, so http2_recv_timeout (replaced by
client_header_timeout in previous patches) is essentially cancelled.
The client_header_timeout is, however, used for SSL handshake and
while reading HEADERS frames.
Instead, keepalive_timeout and keepalive_requests are now used. This
is expected to simplify HTTP/2 code and usage. This also matches
directives used by upstream module for all protocols.
In case of default settings, this effectively changes maximum number
of requests per connection from 1000 to 100. This looks acceptable,
especially given that HTTP/2 code now properly supports lingering close.
Further, this changes default keepalive timeout in HTTP/2 from 300 seconds
to 75 seconds. This also looks acceptable, and larger than PING interval
used by Firefox (network.http.spdy.ping-threshold defaults to 58s),
the only browser to use PINGs.
Instead, the client_header_timeout is now used for HTTP/2 reading.
Further, the timeout is changed to be set once till no further data
left to read, similarly to how client_header_timeout is used in other
places.
New connections are marked reusable by ngx_http_init_connection() if there
are no data available for reading. As a result, if SSL is not used,
ngx_http_v2_init() might be called when the connection is marked reusable.
If a HEADERS frame is immediately available for reading, this resulted
in connection being preserved in reusable state with an active request,
and possibly closed later as if during worker shutdown (that is, after
all active requests were finalized).
Fix is to explicitly mark connections non-reusable in ngx_http_v2_init()
instead of (incorrectly) assuming they are already non-reusable.
Found by Sergey Kandaurov.
If ngx_drain_connections() fails to immediately reuse any connections
and there are no free connections, it now additionally tries to reuse
a connection again. This helps to provide at least one free connection
in case of HTTP/2 with lingering close, where merely trying to reuse
a connection once does not free it, but makes it reusable again,
waiting for lingering close.
This is particularly important in HTTP/2, where keepalive connections
are closed with lingering. Before the patch, reusing a keepalive HTTP/2
connection resulted in the connection waiting for lingering close to
remain in the reusable connections queue, preventing ngx_drain_connections()
from closing additional connections.
The patch fixes it by marking the connection reusable again, and so
moving it in the reusable connections queue. Further, it makes actually
possible to reuse such connections if needed.
This part somehow slipped away from c5840ca2063d.
While it is not expected to be needed in case of lingering close,
it is good to keep it for correctness (see 2b5528023f6b).
Keeping post_accept_timeout in ngx_listening_t is no longer needed since
we've switched to 1 second timeout for deferred accept in 5541:fdb67cfc957d.
Further, using it in HTTP code can result in client_header_timeout being
used from an incorrect server block, notably if address-specific virtual
servers are used along with a wildcard listening socket, or if we've switched
to a different server block based on SNI in SSL handshake.
The stub status module and ngx_http_send_response() (used by the empty gif
module and the "return" directive) incorrectly assumed that responding
to HEAD requests always results in r->header_only being set. This is not
true, and results in incorrect behaviour, for example, in the following
configuration:
location / {
image_filter size;
return 200 test;
}
Fix is to remove this incorrect micro-optimization from both stub status
module and ngx_http_send_response().
Reported by Chris Newton.
